I've long been frustrated by the literalness with which fields like computer science and computer vision approach the topic of seeing. It's a neat trick to point a camera at a picture and have the caption "hot dog" or "not hot dog" appear on the screen. But if you show Rene Magritte's iconic "The Treachery of Images" to such an object recognition model, the classifier will invariably return the result: "This is a pipe." Something is wrong here.
Visual perception is squishy and slippery, formed by each of our unique biological makeups, our memories, history, culture, and our own subjectivities. With the "evolved hallucinations" project, I wanted to see what would happen if we tried to build computer vision models based on a wide range of historical, cultural, and notional worldviews.
